Towards AI-Eiffel Development - Simple Dev

[Liberty Lover]

Good morning, All —

Please find the attached Project Plan document and tell me what you all think.

I am working this out into a detailed design document.

The bottom line is this: While Claude code has been amazing, it has some deep limitations that prevent AI from working easily alongside a human developer or development team. One thing is for sure: Eiffel as both language and method has proven itself in this environment in ways I could not have anticipated, but am delighted to have stumbled into. Nevertheless, the AI+Eiffel+Human model as it stands now is severely limited in quantifiable and qualitative ways. This simple_dev project is going to try and address those limits in an Eiffel-centric environment and context. I think Eiffel is the best to take advantage of AI and the simple_dev is going to work towards proving that by producing a tool that ALL of you can use regardless of which AI you choose to bring along for the ride.

Your thought AFTER reading the document are encouraged and welcomed!!! 

 

I will honestly take ALL feedback from anyone in this forum.

Cheers,

Larry

[Liberty Lover]

You can find the MD version here.

--
You received this message because you are subscribed to the Google Groups "Eiffel Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to eiffel-users...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/1c1087da-d322-46cc-9b27-d70bddc63f5bn%40googlegroups.com.

[Ulrich Windl]

Hi!

Not a comment you expect, but you might want to read anyway:
https://www.heise.de/en/news/39C3-Security-researcher-hijacks-AI-coding-assistants-with-prompt-injection-11125687.html

So better not tell Claude to ruin your computer;-)

Regards,
Ulrich

30.12.2025 18:34:00 Liberty Lover <rix....@gmail.com>:

> --
> You received this message because you are subscribed to the Google Groups "Eiffel Users" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to eiffel-users...@googlegroups.com.

> To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/1c1087da-d322-46cc-9b27-d70bddc63f5bn%40googlegroups.com[https://groups.google.com/d/msgid/eiffel-users/1c1087da-d322-46cc-9b27-d70bddc63f5bn%40googlegroups.com?utm_medium=email&utm_source=footer].

[Liberty Lover]

I will try my hardest to avoid this. :-) 

To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/8adfb968-9914-4bb8-9323-4f4709b7a249%40gmail.com.

[Anders Persson]

Hi Larry

After following your work during the last month, without doing any tests yet with the new libraries, and reading the Project Plan for simple_dev I will try to give you some feedback on it which I hope will get the discussions to take off on this topic.

As I understand the Project Plan the main problem is that you struggled with keeping the Claude AI to learn, when you did your development, and it was frustrating that it did forget a lot during the process. You invented different ways to reinforce how you wanted it to work but it did not get to the next level and it was still a newbie. It also cost a lot to reinforce it.

In the Project Plan you want to wrap the AI tool which would automate the reinforcement of the knowledge you have added to the AI when needed. It should also guide the process of software engineering by enforcing it to do design, test etc as you want it to do work.

I would suggest that we consider building a LLM specialist in Eiffel Software Engineering built on top of a pre-trained model for example Metas LLaMA 4. Something similar to what Code GPT has done. 

Vänligen

Anders Persson

BSharp AB

Linked in profile

+46 763 17 23 25

--

You received this message because you are subscribed to the Google Groups "Eiffel Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to eiffel-users...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/1c1087da-d322-46cc-9b27-d70bddc63f5bn%40googlegroups.com.

[Liberty Lover]

Hi Anders,

A specialized LLM was the ultimate goal, for sure. It may well turn out that this is optimal. Perhaps better and smarter brains that I could put that together. I am a rank amatuer by any stretch of the imagination.

Best,

Larry

To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/CAGh7wYGku71AwXHnbFNfp3P2QEkjFGso5U54ixx%2BU4Vrp41iNg%40mail.gmail.com.

[Anders Persson]

Hi Larry

I have now made some investigations and tested how it works with building my own AI model for Eiffel. I have defined the following 3 steps in my project.

1. Evaluation of Llama: Set up a pipeline that verifies that I can teach a pre-built model Eiffel specific knowledge.

2. Proof of Concept: Define a set of rules to teach the model Eiffel Engineering and verify the approach.

3. Production: Invest in HW to build the production version of it.

The first step has been completed and I have now started with step 2. There are 3 types of Instruction response pairs.

a. Correct Eiffel

b. Process oriented

c. Critique and correction

I have created a github repository where I add these. It is a private repository but feel free to join by sending me an email. Anyone that contributes to creating these pairs, minimum 200, but 500 is the goal, will get the final version of the model.

Since the investment in HW will be $10k+ I want to keep it within a small set of people. 

Vänligen

Anders Persson

BSharp AB

Linked in profile

+46 763 17 23 25

To view this discussion visit https://groups.google.com/d/msgid/eiffel-users/CA%2B3qnjevnCB702Rdfkki%2BhAcd%3D46_tX8Z%2Bf8u4M2HtiFfuWdEg%40mail.gmail.com.